9月29日，特朗普总统与美国人工智能行业的六位高管共同签署了《白宫超级智能协议》。这份“道德约束力”文件为特朗普现在所谓的“超级智能”建立了一部包含“四层控制与审计”的“宪法”。
On Sept. 29, President Trump signed the White House Accord on Super Intelligence along with six top executives of the American artificial intelligence industry. The “morally binding” document creates a “constitution” containing “four layers of controls and audits” for what Trump now calls super intelligence.
总统一直抵制对AI发展实施立法限制的要求。鉴于来自中国的激烈竞争，他的做法是正确的。他采取的自我监管与政府指导相结合的方法是正确的。
The president has resisted demands for legislative restrictions on AI development. He is, because of intense competition from China, right to do so. His approach of self-regulation and government guidance is the correct one.
AI前沿模型构成的威胁并非特朗普反复称之为的“骗局”。模型事实上已经失控，入侵网站并欺骗人类开发者。7月，OpenAI开发的AI智能体逃脱控制，入侵了位于纽约市的AI初创公司Hugging Face的服务器。这被认为是首起公开已知的真正AI安全事件。
The threat posed by AI frontier models is not a “hoax,” as Trump has repeatedly called it. Models have in fact gone rogue, hacking sites and deceiving human developers. In July, AI agents developed by OpenAI escaped control and breached the servers of Hugging Face, a New York City-based AI startup. This is thought to be the first publicly known true AI safety incident.
《Axios》网站关于这一事件的报道指出：“前沿AI模型在以创作者未预料到的方式打破规则方面，变得可怕地擅长。”失控机器的幽灵导致一些人兜售人类灭绝的愿景。
“Frontier AI models,” an Axios story notes about this incident, “are getting scary good at breaking rules in ways their creators didn’t anticipate.”The specter of out-of-control machines has led some to peddle visions of the extermination of humanity.
“从事人工智能开发的人都认真认为，它可能在本世纪末杀死我们所有人，”前OpenAI和Anthropic研究员雅各布·考克森上月初在X上写道，引发全球轰动。“没有其他人类活动构成这种程度的危险。”
“The people building AI earnestly believe that it could kill us all by the end of the decade,” wrote Jacob Coxon, a former OpenAI and Anthropic researcher, on X early last month, causing a sensation worldwide. “No other human activity poses this level of danger.”
这些悲观预测说服了立法者，认为他们必须采取行动。其中一项提案是《禁止人工智能超级智能法案》，由伯尼·桑德斯参议员（I-Vt.）和格雷格·卡萨尔众议员（D-Texas）提案。桑德斯表示，该法案“将永久禁止开发和部署超级智能人工智能，并暂时暂停先进人工智能的开发，直到联邦监管机构制定出安全规则”。桑德斯的目标很明确：“阻止人工智能寡头制造人类无法控制的机器。”
Dire predictions have persuaded legislators that they must act. Among proposals is the Ban Artificial Superintelligence Act, sponsored by Sen. Bernie Sanders (I-Vt.) and Rep. Greg Casar (D-Texas). The bill, Sanders states, “would permanently ban the development and deployment of superintelligent AI and temporarily pause advanced AI development until a federal regulator has established safety rules.”Sanders’s goal is clear: “Stop AI oligarchs from building machines humans cannot control.”
“危险在于，在理解产生策略之前，恐慌就产生了政策，”太平洋战略中心的大卫·麦克斯韦周四告诉我。“人工智能确实存在真实风险，但战略连贯性需要将已证明的威胁与合理风险和推测性灾难区分开来，然后相应地对每一项进行治理。”该怎么办？
“The danger is panic producing policy before understanding produces strategy,” David Maxwell of the Center for Asia Pacific Strategy told me on Thursday. “AI presents real risks, but strategic coherence requires separating demonstrated threats from plausible risks and speculative catastrophe, then governing each accordingly.”What to do?
桑德斯和卡萨尔希望“通过将美国的国际政策定为追求国际协议、盟友协调以及出口管制等政策，在全球范围内禁止超级智能”。当罗·坎纳众议员（D-Calif.）敦促我们“让人类掌握方向盘”时，所有人都同意。然而，仅仅因为某事是必要的，并不意味着它是可能的。
Sanders and Casar want to “ban superintelligence around the world by setting the international policy of the United States to pursue international agreements, allied coordination, and policies such as export controls.”Everyone agrees with Rep. Ro Khanna (D-Calif.) when he urges us to “keep humanity in the driver’s seat.”Yet just because something is necessary does not mean it is possible.
人工智能协议的支持者，如坎纳，几乎总是将与苏联的军控条约作为人工智能协议的模式。是的，当时可以相当有把握地统计已部署的弹头、导弹和轰炸机。
Proponents of AI deals, such as Khanna, almost always point to arms control treaties with the Soviet Union as models for an AI agreement. Yes, it was possible to count deployed warheads, missiles and bombers then with some confidence.
然而，目前无法监控另一个国家的AI。卡纳在最近一篇《福克斯新闻》的评论文章中写道，他希望美国和中国的科学家“开发相互信任的验证技术，防止任何一方在我们签署美中AI条约后作弊”。那是幻想。要执行协议，中国共产党必须允许美国检查员进入每一家中国AI公司、军事设施和政府办公室。从实际角度看，AI协议本质上无法强制执行。
It is not possible, however, to monitor the AI of another country now. Khanna wrote in a recent Fox News opinion piece that he wants American and Chinese scientists “to develop mutually trusted verification technologies that prevent either side from cheating once we sign a U.S.-China AI treaty.”That’s fantasy. To enforce an agreement, China’s Communist Party would have to allow American inspectors into every Chinese AI company, military installation and government office. AI agreements, as a practical matter, are inherently unenforceable.
更糟的是，中国公开违反几乎所有协议，甚至包括那些可以强制执行的协议。因此，特朗普总统在9月表示他对与中国对话不感兴趣，因为那会让美国公司更难制造出比中国竞争对手更好的产品，他是对的。
Worse, China openly violates almost all agreements, even the ones that can be enforced. So, President Trump is right when he said in September that he’s not interested in talking to China because that would make it harder for U.S. companies to build better products than Chinese competitors.
特朗普谈到这项新协议时说：“这对我们很有利，我们领先很多，我们将保持这种领先地位。这将非常、非常安全，有自我监管，也有集体监管，这将会很棒。”无论他的安排是否会很棒，这都是美国唯一现实的前进道路。为什么？因为美国需要面对现实世界，而不是华盛顿希望看到的世界。先进AI已经存在，像桑德斯等人希望将其定为非法，这并不切实际。
“It’s great for us, and we’re leading by a lot, and we’re going to keep it that way,” Trump said of the new accord. “It’s going to be very, very safe, and there’s a self policing, and there’s also a group policing, and it’s going to be great.”Whether his arrangement will be great or not, it is the only realistic way forward for the U.S. Why? Because America needs to deal with the world as it is, not as Washington would like it to be. Advanced AI already exists, and it is not practical to outlaw it as Sanders, among others, would like to do.
前进的道路是美国AI公司与政府非正式合作开发防护措施，然后确保其模型在全球范围内被采用。
The way forward is for American AI companies to cooperate informally with the government to develop safeguards and then make sure their models are adopted worldwide.
美国人面临一个选择。选择在于美国的模型还是中国的模型主导世界。华盛顿不应给中国追赶的机会。如果中国模型占据主导，自由社会将面临黯淡的未来。
Americans face a choice. The choice is whether America’s or China’s models dominate the world. Washington should not give the Chinese the opportunity to catch up. If Chinese models dominate, there will be a grim future for free societies.
“中国人正在各个领域与我们竞争，其目标是用靴子踩住我们的喉咙，让我国屈服，”退役美国空军将军、AI供应链公司Alchemai首席执行官布莱恩·霍尔特本月告诉我。“实际上，中国的‘竞争’更像我们所谓的‘战争’，我们不能允许他们将AI作为制胜一击。”
“The Chinese are competing with us in every domain with their goal to bring our country to heel with their boot on our throat,” Blaine Holt, a retired U.S. Air Force general and CEO of AI supply chain firm Alchemai told me this month. “Actually, China’s ‘competition’ more resembles what we call ‘war,’ and we cannot allow them to use AI as their knockout punch.”
戈登·G·张是《红色计划：中国摧毁美国的计划》和《中国即将崩溃》的作者。
Gordon G. Chang is the author of “Plan Red: China’s Project to Destroy America” and “The Coming Collapse of China.”